🎲 Ideogram 4 Random Prompter

Ideogram 4 only really works if you speak its JSON schema - hex palettes, bounding boxes, typed regions. Writing that by hand, every time, is an hour of your life you don't get back. 🎲 Ideogram 4 Random Prompter skips the authoring entirely: it rolls a whole structured caption for you, boxes and all, and hands you the JSON.

It's marked experimental, and the author means it - this is a generator, not a precision tool. But it's genuinely fun and it does a useful job: you get to see what the schema looks like when it's valid, then either keep the roll or hand-edit the good parts in the 🧩 Ideogram 4 Prompt Builder. Treat it as a slot machine that teaches you the format.

How the randomness works

Every word is pulled live from the wonderwords dictionary - the node's description makes a point of it, nothing is hardcoded - so the vocabulary is broad and English. On top of that it scatters regions across the canvas in four size tiers with independent weights: background_weight, large_weight, medium_weight, small_weight. All four get summed and each region picks a tier proportionally, so setting all of them to 0 falls back to equal weighting, and setting one to 1.0 with the rest at 0 forces the whole composition to one scale. A background block covers at least 70% of the canvas area - it's a base layer. Small elements are ~5-17% per axis.

region_count_min / region_count_max decide how many regions, and the count is a random integer in that range inclusive. Set them equal for a fixed count.

Then the parts that change the character of the output more than any setting:

• scene_framing - on, the dictionary words get woven with articles and spatial connectors into one continuous sentence ("a vivid tower beside a hollow stone against an amber cloud"), which tells the model it's a single coherent scene. Off, you get bare word lists per region, and Ideogram will happily render that as a collage or asset sheet of separate objects. If your "photograph" keeps coming out looking like a grid, this is why.
• freeform_chance - the per-region probability that a region drops its bounding box entirely. Ideogram has no edge-blur control, so an element without a box blends softly into the scene instead of being pinned to a hard rectangle. Turn it up to escape the cut-out look; at 1.0 nothing is boxed and the bboxes output is empty.
• text_region_min / text_region_max - an exact count of regions rendered as in-image text rather than objects. 0/0 for never, 3/3 for always exactly three words in the picture.

medium defaults to photograph (which emits a photo style key with focal length and aperture; anything else emits an art_style key). color_palette mirrors the Colorful Starting Image node's palettes - none emits no palette at all and lets Ideogram choose freely. color_harmony, positioning_bias and arrangement control how colours and centres are distributed. word_length_bias and word_length_randomness bias the dictionary words toward a character length, which is a weirdly effective way to shift tone from punchy to ornate.

description_length, description_override, description_prefix and description_background_prefix shape the auto-generated overview and background lines; description_override wins outright and turns off generation for that field.

Outputs

Same shape as the builder: prompt (the caption JSON), preview (rendered regions), bboxes (pixel boxes as {x, y, width, height}, freeform regions excluded), width, height. seed makes any roll reproducible - same seed plus same settings gives you the identical caption, preview and boxes, which is how you go back to the one that worked.

The one thing to try first

The tooltip practically hands you a recipe: for something that reads as an actual photo, use scene_framing on, medium: photograph, a low region count, and high background/large weights. High region counts with tiny boxes and freeform off is how you get a busy collage - which is a great look, just not a photo.

And keep an eye on the token budget downstream. Ideogram 4 caps its caption at 2048 tokens; long descriptions across 25 regions is how you find that ceiling.

Install

ComfyUI Manager → ComfyUI-mnemic-nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MNeMoNiCuZ/ComfyUI-mnemic-nodes

Restart. This node needs the wonderwords package, which is in the pack's requirements.txt - Manager handles it, but a manual git clone without reinstalling requirements leaves you with a node that fails on the dictionary import. If it won't load, that's what to check first:

python -m pip install wonderwords

You'll also need an Ideogram 4 checkpoint and a JSON-capable prompt path for the output to be good for anything. Remember the model is non-commercial licensed - fine for experiments, worth reading the agreement before you build a business on it.